AI integration & agents
Put AI to work inside your business — properly.
A chatbot on your website is the start, not the limit. Real AI integration means wiring AI into how your business actually works: answering from your own documents instead of guessing, agents that carry out whole tasks in the tools you already use, classifiers that sort and route work automatically, and — where it matters — AI running privately so your data never leaves your control. I bring the engineering side of this: grounding answers in your data (RAG), building agents with safe access to your systems (MCP, tools, skills), tuning models to your specific work, and setting up local AI where privacy or cost demands it. The point isn't to look clever with AI — it's to remove real work and reduce real mistakes.
When this service makes sense
You probably need this if…
You want AI to answer from your own documents, prices, or policies — accurately — not make things up.
You'd like an assistant that can actually do things in your tools (your calendar, CRM, spreadsheets), not just talk.
A chunk of your day is sorting things by hand — enquiries, emails, invoices, documents — that a machine should be routing for you.
Off-the-shelf AI gives generic answers, and you need it tuned to how your business really works.
Your data is sensitive and you'd rather AI ran privately, on your own kit, than sent off to a third party.
You've tried an AI tool and it's inconsistent, expensive, or doesn't fit your workflow.
How I approach it
My approach, step by step.
- 01
Ground it in your own knowledge (RAG)
Rather than hope a model 'knows' your business, I connect it to your real documents — handbooks, price lists, past quotes, policies — so answers are pulled from your actual information and can be traced back to the source. This is retrieval-augmented generation (RAG), and it's the difference between a confident guess and a correct answer.
- 02
Get more from prompting before anything heavier
A lot of 'we need a custom model' turns out to be a prompting problem. With well-chosen examples (few-shot prompting) and clear instructions, an off-the-shelf model often does the job — faster and cheaper than training. I start here and only escalate when the results genuinely need it.
- 03
Connect it to your real tools (MCP)
Using the Model Context Protocol (MCP), I give an assistant safe, controlled access to the systems you already use — so it can check availability, draft a reply, update a record, or pull a number, instead of leaving you to do the copy-paste. Permissions and guardrails are part of the build, not an afterthought.
- 04
Build agents that do whole jobs, not just answer
An agent goes further than an assistant: given a goal — chase this invoice, prepare this quote, triage this inbox — it works through the steps itself, using the tools and skills I've built for it, and checks in with you at the points you decide. I build agents with clear boundaries, so they're genuinely useful without ever acting beyond what you've approved.
- 05
Sort and route automatically (AI classifiers)
A lot of admin is just deciding which pile something goes in — is this enquiry a lead or spam, which job does this invoice belong to, is this review worth replying to. I build classifiers that make those calls automatically and reliably, measured against real examples of your work, so the routine sorting happens without you.
- 06
Tune a model to your work — when it pays off
When prompting isn't enough, I fine-tune a model on examples of your work so it adopts your tone, format, and judgement. I also set up the quieter half of this: efficient orchestration to collect and label the right data from your day-to-day operations, so a fine-tune is built on solid examples rather than guesswork.
- 07
Run it privately when it matters
For sensitive data or predictable costs, I can configure AI to run locally — on your own hardware or a private server — so nothing is sent to an outside provider. You get the capability without handing your data over, and without a per-message bill that climbs with usage.
What you get
Concrete deliverables.
- AI grounded in your own documents (RAG), with answers you can trust and trace
- Agents connected to your real tools via MCP — with the tools and skills to act, not just chat
- AI classifiers that sort and route enquiries, emails, and documents automatically
- Prompting set up properly — few-shot examples and guardrails — before any heavier build
- Fine-tuned models where they pay off, plus the data-collection pipeline to feed them
- Local / private AI configuration where privacy or cost makes it the right call
- Plain-English documentation and a handover so you understand what's running
Typical timeline
A grounded assistant (RAG) over your documents can come together quickly — often a week or two. Fine-tuning, custom tool connections (MCP), and local AI setups take longer because there's data and infrastructure involved. I scope it in a free call and give you a clear plan and fixed-or-day-rate quote before any work starts.
Common questions
What clients usually ask.
What's the difference between this and an AI chatbot?
A chatbot is one thing AI can do — answer visitors on your site. AI integration is the broader engineering: grounding answers in your own data (RAG), letting AI take actions in your tools (MCP), tuning models to your work, and running AI privately where needed. If you just want a website assistant, start with the chatbot service; if you want AI woven into how the business runs, this is it.
What's an AI agent, and how is it different from a chatbot?
A chatbot answers questions; an agent gets things done. You give it a goal — chase the unpaid invoices, triage this morning's enquiries — and it works through the steps itself using controlled access to your tools, checking in with you where you've told it to. Building one well is an engineering job: the connections (MCP), the tools and skills it uses, and the guardrails around it all have to be done properly. That's the part I bring.
What is RAG, in plain English?
Retrieval-augmented generation. Instead of relying on what a model was trained on, it looks up the relevant bits of your own documents first, then answers from those. That's how you get accurate, up-to-date answers about your business — and how you avoid the confident-but-wrong replies AI is known for.
Do I really need a fine-tuned model?
Often not — and I'll tell you honestly. Good prompting with the right examples (few-shot) and grounding in your data (RAG) solves most needs faster and cheaper. Fine-tuning is worth it when you need a consistent tone or format at scale, and only once there's good data to train on — which I can help collect and organise.
Can the AI run privately, without sending my data to a third party?
Yes. Where privacy or cost matters, I can set up AI to run locally — on your own hardware or a private server — so your data stays with you. There's usually a trade-off between capability, cost, and privacy, and I'll lay it out clearly so you can choose with open eyes.
What does AI integration cost?
It depends on scope — a grounded (RAG) assistant over your documents is a smaller piece than fine-tuning a model or standing up local AI infrastructure. Simpler builds start from around £500; deeper work is quoted at a day rate. There can also be a small per-use AI cost from the provider, unless we run it locally. I'll give you a clear, itemised quote after a free call.
Related services
Often paired with this.
Want to talk about ai integration & agents?
Tell me what you're trying to do. It's me who reads your message, and I'll reply within one business day.
No deck. No drip campaign. One reply.